Federated Learning With Energy Harvesting Devices

Federated learning (FL) is a promising technique for distilling artificial intelligence from massive data distributed in Internet-of-Things networks, while keeping data privacy. However, the efficient deployment of FL faces several challenges due to, e.g., limited radio resources, computation capabilities, and battery lives of Internet-of-Things devices. To address these challenges, in this work, the energy harvesting technique is first enabled on Internet-of-Things devices for supporting their sustainable lifelong learning. Then, the convergence rate of the FL algorithm is derived, which is shown to depend on the data utility (defined as the number of used training samples) in each training iteration. Thus, to accelerate the convergence rate and reduce the training latency, a data utility maximization problem for each iteration is formulated, under several practical constraints on the limited time, bandwidth (i.e., number of subcarriers), computation frequency, and energy supply. The problem is mixed-integer and non-convex, and hence NP-hard. To solve the problem, an optimal joint device selection and resource allocation (JDSRA) scheme is proposed. In this scheme, a distributed on-device resource allocation problem is first solved to determine the minimum required number of subcarriers for each device, followed by a dynamic programming approach for attaining the optimal device selection policy. In particular, no global channel state information (CSI) sharing is needed to execute the scheme. Finally, extensive experiments are presented to demonstrate the performance of the proposed optimal algorithm.